SSML for AI Narration: Control Your Voice

Q: Which SSML tags control pauses and emphasis?

Use the tag for pauses, setting either a duration like or a strength like . Use the tag with a level of reduced, moderate, or strong to stress specific words. For broader pacing, wrap text in and set the rate, pitch, and volume attributes.

Q: How do I fix a mispronounced word or brand name?

Wrap it in a pronunciation tag. Use to substitute a spoken form for written shorthand, to control how numbers and dates are read, and to spell out the exact sounds of a tricky proper noun. Tagging names and acronyms by default is the fastest way to avoid rerolls.

SSML for narration is the difference between an AI voice that reads your script and one that performs it. Speech Synthesis Markup Language lets you control pauses, emphasis, pacing, and pronunciation tag by tag, so a voiceover lands the way you would direct a human narrator. This guide explains what SSML is, the specific tags that shape delivery, and how to use pacing controls inside CoreReflex's engine-agnostic voice seam.

What SSML is and why it matters for AI narration

SSML is a standard XML-based markup language for text-to-speech. Instead of feeding a synthesizer a flat string of words and hoping for the best, you wrap that text in tags that tell the engine how to say it, where to breathe, which word to stress, how fast to move, and how to pronounce anything ambiguous. The plain text is the script; the SSML is the direction.

For narration specifically, this matters enormously. A documentary read needs unhurried pauses; a product explainer needs crisp pacing and a stressed benefit; an audiobook needs sentence rhythm that doesn't fatigue the ear. Raw text-to-speech treats all of these the same. SSML is how you give each one its own cadence. If you have ever felt that an AI voiceover sounded almost right but mechanically flat, the missing ingredient is usually direction, and that is exactly what making AI voiceover sound natural comes down to.

The SSML tags that control pacing and emphasis

Most of the expressive power you need comes from a small set of tags. Where the underlying voice engine supports SSML. These are the ones that earn their keep.

Pauses with the break tag

The <break> tag inserts silence, and silence is the single most underused tool in narration. A pause before a key phrase creates anticipation; a pause after it lets the point land. You control duration explicitly, for example <break time="700ms"/> for a beat or <break time="1.5s"/> for a full dramatic stop. You can also use strength values like <break strength="strong"/> to mimic the natural pause at a sentence boundary. Good narration is often defined by where the silences fall, not just the words.

Emphasis

The <emphasis> tag stresses a word or phrase, with levels typically ranging from reduced to moderate to strong. Wrapping a single word, <emphasis level="strong">free</emphasis>, pushes the engine to lean on it the way a human would when a word carries the meaning of the sentence. Use it sparingly; emphasis everywhere reads as emphasis nowhere.

Prosody: rate, pitch, and volume

The <prosody> tag is your master pacing control. It exposes three attributes:

rate, speed, set with keywords like slow and fast or percentages like rate="90%" to slow a dense technical passage.
pitch, vocal height, raised or lowered to shift energy and tone.
volume, loudness, useful for pulling a line back to a near-whisper or pushing it forward.

Wrapping a paragraph in <prosody rate="95%">…</prosody> is often all it takes to turn a rushed read into a composed one. Pacing is the lever that most changes how trustworthy a narration feels, which is why pauses, speed, and emphasis deserve their own deeper treatment in our guide to AI voiceover pacing.

Pronunciation: say-as, sub, and phoneme

The tags that save you the most rerolls are the pronunciation controls.

<say-as interpret-as="date"> or interpret-as="telephone" tells the engine how to read numbers, dates, and codes so "1999" becomes "nineteen ninety-nine" rather than "one thousand nine hundred ninety-nine."
<sub alias="World Health Organization">WHO</sub> substitutes a spoken form for written shorthand.
<phoneme> lets you spell out the exact pronunciation of a tricky brand name or proper noun.

For any narration with product names, acronyms, prices, or dates, these tags are the fastest path to a clean take.

A practical SSML workflow in CoreReflex

Inside CoreReflex, voiceover runs through an engine-agnostic voice seam, the same seam that powers narration for films, script reads, and the voice that answers the phone. Here is a workflow that produces a polished narration track without trial-and-error reloading.

Write the script in plain text first. Get the words right before you touch a single tag. Direction can't fix a weak script.
Choose your voice. Pick a named HD voice, a voice you cloned with consent, or one you created in Design a Voice mode. The voice sets the baseline character; SSML shapes the performance on top of it.
Mark the breaths. Add <break> tags where a human narrator would pause, before a turn, after a claim, at the end of a thought. This single pass does the most for naturalness.
Stress the meaning. Add <emphasis> to the one or two words per sentence that carry the point. Read it aloud yourself first to find them.
Set the pace. Wrap dense or fast sections in <prosody rate="…"> to slow them, and leave conversational sections at default.
Lock pronunciations. Tag every name, acronym, number, and date with <say-as>, <sub>, or <phoneme> so the engine never guesses.
Generate and place. Render the track and drop it on the timeline, where it sits alongside your shots, music, and the deterministic render path that produces the same cut from the same manifest every time.

Because planning and editing are free in CoreReflex and only generation consumes credits, you can refine your tags and re-listen without watching a meter, the iteration loop on a narration track is meant to be cheap.

Common SSML mistakes to avoid

A few habits separate clean SSML from markup that fights the engine:

Over-tagging. Emphasizing every other word flattens the read. Direct the meaning, not the whole sentence.
Unclosed tags. SSML is XML, every <prosody> needs its </prosody>. A single unclosed tag can break the whole block.
Fighting the voice. If a voice's natural pace is wrong for the piece, change the voice before you bury it in prosody overrides. The right starting voice does most of the work, which is why choosing the right AI voice is upstream of any tagging.
Ignoring numbers. Untagged numbers, dates, and currency are the number-one cause of awkward reads. Tag them by default.

SSML across the whole voice seam

The payoff of learning SSML is that the same direction follows your brand voice everywhere. The voice seam is engine-agnostic by design, so the pacing and pronunciation choices you make for a film narration carry over to script reads and to long-form work like audiobook and long narration projects. The same brand voice can even answer the phone through real-time voice agents and appointment setters on Gemini Live, and the pronunciation rules you defined for your product names hold there too. You direct the voice once; it stays in character across every surface. The full set of supported controls is documented in the voice docs, and you can hear the difference in the Voice pillar.

Frequently asked questions

What is SSML and why use it for voiceover?

SSML, or Speech Synthesis Markup Language, is a standard XML markup that controls how a text-to-speech engine speaks, pauses, emphasis, pacing, pitch, volume, and pronunciation. You use it for voiceover because plain text gives the synthesizer no direction, so it reads everything with the same flat cadence. SSML lets you direct the performance the way you would direct a human narrator, which is what makes AI narration sound intentional rather than robotic.

Which SSML tags control pauses and emphasis?

Use the <break> tag for pauses, setting either a duration like <break time="700ms"/> or a strength like <break strength="strong"/>. Use the <emphasis> tag with a level of reduced, moderate, or strong to stress specific words. For broader pacing, wrap text in <prosody> and set the rate, pitch, and volume attributes.

Does every AI voice support SSML?

SSML support depends on the underlying engine, which is why CoreReflex uses an engine-agnostic voice seam rather than tying you to one synthesizer. Where the engine behind your chosen voice supports SSML, the pacing and pronunciation tags apply; the pacing controls exposed on the seam give you direction over the read regardless. Check the voice docs for the controls available to the specific voice you select.

How do I fix a mispronounced word or brand name?

Wrap it in a pronunciation tag. Use <sub alias="…"> to substitute a spoken form for written shorthand, <say-as> to control how numbers and dates are read, and <phoneme> to spell out the exact sounds of a tricky proper noun. Tagging names and acronyms by default is the fastest way to avoid rerolls.

Direct the voice, don't just generate it

SSML turns AI narration from a flat read into a directed performance, pauses where they belong, stress on the words that matter, pacing tuned to the piece, and pronunciations locked down. In CoreReflex, those controls live on an engine-agnostic voice seam that carries your brand voice from films to scripts to the phone, all feeding a deterministic render path that reproduces the same cut every time. Pick a voice, mark up a script, and hear the difference. It is free to start, no credit card.